Agentic employees in world's most downloaded FinTech app

Imagine 20 engineering teams, each with their own AI engineering teammate—reviewing PRs, triaging incidents, automating repetitive tasks 24/7, answering on daily support tickets, investigating reports, creating action points and improving based on the provided feedback.

We scaled AI-powered automation from a single proof-of-concept to autonomous agents across our organization, each acting as a virtual engineer for their team. Unlike centralized chatbots that provide generic answers, each team’s AI has deep domain expertise: it knows their codebase, their Jira workflows, their on-call runbooks. These bots don’t just suggest solutions—they execute them. Collectively, they handle 300+ hours of work per week, automating 55-65% of routine platform engineering tasks.

But scaling AI automation isn’t just multiplying one bot by 20. We faced brutal challenges: How do you prevent credential leakage when teams can trigger bot actions? How do you govern security across 20 independent agents? How do you share learnings across teams without creating a bloated one-size-fits-all monster?

This talk reveals the architecture, economics, and organizational dynamics of running AI teammates at scale. You’ll learn why we chose physical laptops over VMs, how we built a security framework that catches attacks the LLM misses, and the surprising patterns that work across teams (PR reviews, incident triage) versus those that need deep customization (team-specific deployments, domain knowledge).

We’ll share real metrics: 300+ hours/week saved, sub-1% security false positive rate, 60% ticket automation rate, and the feedback loop that improves all 20 bots simultaneously. You’ll see the mental shifts teams went through from “the bot is a toy” to “the bot is on-call,” and why treating AI as an employee (not a service) changed everything.